Low Latency Variational Autoencoder on FPGAs

Variational Autoencoders (VAEs) are at the forefront of generative model research, combining probabilistic theory with neural networks to learn intricate data structures and synthesize complex data. However, designs targeting VAEs are computationally intensive, often involving high latency that precludes real-time operations. This paper introduces a novel low-latency hardware pipeline on FPGAs for fully-stochastic VAE inference. We propose a custom Gaussian sampling layer and a layer-wise tailored pipeline architecture which, for the first time in accelerating VAEs, are optimized through High-Level Synthesis (HLS). Evaluation results show that our VAE design is respectively 82 times and 208 times faster than CPU and GPU implementations. When compared with a state-of-the-art FPGA-based autoencoder design for anomaly detection, our VAE design is 61 times faster with the same model accuracy, which shows that our approach contributes to high performance and low latency FPGA-based VAE systems.